Anchor-free Object Detectors

最新推荐文章于 2022-07-04 21:57:57 发布

dekiang

最新推荐文章于 2022-07-04 21:57:57 发布

阅读量362

点赞数

分类专栏： Object Detection

本文链接：https://blog.csdn.net/weixin_41560402/article/details/109016692

版权

55 篇文章 17 订阅

订阅专栏

1. keypoint-based anchor-free detector

该类目标检测模型主要将目标检测问题看成是关键点检测和分组问题，不同的算法的区别也在于这两个方面，即

CornerNet检测2个keypoint：top-left corner, bottom-right corner；
CenterNet检测3个keypoint：top-left corner, bottom-right corner, center point；
ExtremeNet检测5个keypoint：top-most, leftmost, bottom-most, right-most, center point。
CentripetalNet检测2个keypoint(top-left corner, bottom-right corner)，对预测每个keypoint的中心点；

CornerNet, CenterNet：为每个keypoint生成一个associative embedding，根据associative embedding的相似性进行分组；
ExtremeNet： appearance-based geometric grouping，即预测框的中心在center map上的置信度是否大于某一阈值。
CentripetalNet：根据keypoint所对应的中心点的距离进行关键点分组；

该类目标检测模型更接近于anchor-based模型，即是把anchor box换成anchor point，在特征图的每一位置只输出一个预测结果，如DenseBox, UnitBox, YOLOv1, FCOS, FoveaBox等。
FCOS在center-based anchor-free detector中的作用等同于RetinaNet在one-stage anchor-based detector中的作用，后续许多算法都是在FCOS的基础上进行研究的，并且大多数都是研究如何进行label assignment的，如ATSS, FSAF, AutoAssign等，label assignment经历了从固定、半动态，最后到全动态的优化过程。
Anchor-free detector逐渐代替anchor-based而成为主流的目标检测算法，其实，anchor-free detector是最近2年（2019和2020年）才开始蓬勃发展的，所以anchor-free算法还有很多提升空间，并且其最优性能、模型泛化能力和应用场景将优于anchor-based方法。

CornerNet: Detecting Objects as Paired Keypoints [ECCV] 【CornerNet】
An Anchor-Free Region Proposal Network for Faster R-CNN based Text Detection Approaches 【AF-RPN】

ExtremeNet: Bottom-up Object Detection by Grouping Extreme and Center Points [CVPR] 【ExtremeNet】
Feature Selective Anchor-Free Module for Single-Shot Object Detection [CVPR] 【FSAF】
CenterNet: Keypoint Triplets for Object Detection [CVPR] 【CenterNet】
FCOS: Fully Convolutional One-Stage Object Detection [ICCV] 【FCOS】
RepPoints: Point Set Representation for Object Detection [ICCV]
FoveaBox: Beyond Anchor-based Object Detector 【FoveaBox】
CornerNet-Lite: Efficient Keypoint Based Object Detection
Objects as Points
Segmentation is All You Need.

Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection [CVPR] 【ATSS指出one-stage anchor-based与one-stage anchor-free的性能差异在于正负样本的定义不同，并提出ATSS根据anchor与ground truth之间的数据统计信息来定义正负样本】
Soft Anchor-Point Object Detection [CVPR] 【 SPAD】
CentripetalNet: Pursuing High-quality Keypoint Pairs for Object Detection [CVPR] 【CentripetalNet】
BorderDet: Border Feature for Dense Object Detection [ECCV] 【BorderDet】
Corner Proposal Network for Anchor-free, Two-stage Object Detection [ECCV]
HoughNet: Integrating near and long-range evidence for bottom-up object detection [ECCV]
Dense RepPoints: Representing Visual Objects with Dense Point Sets [ECCV]
AutoAssign: Differentiable Label Assignment for Dense Object Detection 【AutoAssign】
SaccadeNet: A Fast and Accurate Object Detector
Localization Uncertainty Estimation for Anchor-Free Object Detection.
RepPoints V2: Verification Meets Regression for Object Detection
Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection 【Generalized Focal Loss】
Reducing Label Noise in Anchor-Free Object Detection
CircleNet: Anchor-free Detection with Circle Representation

关注